On 29.09.2011 1:20, Nick Sabalausky wrote:
"Jonathan M Davis"<[email protected]> wrote in message
news:[email protected]...
On Wednesday, September 28, 2011 13:43 Nick Sabalausky wrote:
"Jonathan M Davis"<[email protected]> wrote in message
news:[email protected]...
I would point out that there is an intention to eventually get a D
lexer
and
parser into Phobos so that tools can take advantage of them. Those
could
eventually lead to a frontend in D but would provide benefits far
beyond
simply
having the compiler in D.
Is the interest more in a D-specific lexer/parser or a generalized one?
Or
is it more of a split vote? I seem to remember interest both ways, but I
don't know whether there's any consensus among the DMD/Phobos crew.
A generalized lexer is nothing more than a regex engine that has more
than
one distinct accept state (which then gets run over and over until EOF).
And the FSM is made simply by doing a combined regex "(regexForToken1 |
regexForToken2 | regexForToken3 | ... )", and then each of those parts
just get their own accept state. Which makes me wonder...
There was a GSoC project to overhaul Phobos's regex engine, wasn't there?
Is that done? Is it designed in a way that the stuff above wouldn't be
real hard to add?
And what about algoritm? Is it a Thompson NFA, ie, it traverses the NFA
as
if it were a DFA, effectively "creating" the DFA on-the-fly)? Or does it
just traverse the NFA as an NFA? Or does it create an actual DFA and
traverse that? An actual DFA would probably be best for a lexer. If a
DFA,
is it an optimized DFA? In my (limited) tests, it didn't seem like
DFA-optimization would yield a notable benefit on typical
programming-langauge tokens. It seems to be more suited to pathological
cases.
There is some desire to have a lexer and parser in Phobos which basically
have
the same implementation as dmd (only in D instead of C++). That way,
they're
very close to the actual compiler, and it's easy to port fixes and
improvements between the two.
However, we definitely also want a more general lexer/parser generator
which
takes advantage of D's metaprogramming capabalities. Andrei was pushing
more
for that and doesn't really like the idea of the other, since it would
reduce
the desire to produce the more general solution. So, this _is_ some
dissension
on the matter. But there's definitely room for both. It's just a question
of
time and manpower.
Boy, I gotta say I'm really tempted to tackle this. I don't know if I
*should* dedicate my already-tight time, but it's very tempting. And I have
already written a generalized lexer generator in D (
www.semitwist.com/goldie ), so I have that experience (and codebase) to draw
upon.
Interesting and I almost forgot that we have lexer generator... What
that "generalized" bit applies to? Does it tackle CFG? Then that would
have been parser in my vocabulary ;)
Judging by first pages I see LALR(1) so definitely a parser.
I'm more into LL+something or PEGs. I'm liking the way e.g. ANTLR does
this, a very nice hybrid approach.
Only big question is whether it would be best to try to make Phobos's
existing regex engine flexible enough that it could be used by the lexer
(since a generalized lexer is essentially a regex engine with multiple
accept states, and optionally some customizable hooks). I've posted some
questions to that end in another branch of this thread.
To that end all what needs to be done is to restrict some wild stuff
like backreferences & lookaround (how the hell thought that was good
idea?!). Then use existing parser to get IR code for regex per each
alternative, then fuse them via thompson construction (keeping note of
terminal states).
Taking Unicode into account I'd rather not go for table driven DFA. I'd
better craft some switch statements and let the compiler sweat :)
--
Dmitry Olshansky
